Cerebras Systems is behind schedule on the 750 megawatts of inference capacity OpenAI has committed to buy, and the delay is already costing it credits. Chief executive Andrew Feldman told analysts on the third-quarter earnings call that “we find ourselves running behind schedule on initial tranches”, and the Form 10-Q filed on 12 November says service credits “commenced accruing as of October 1, 2026”.
Less than 5% of the committed capacity could become cancellable this year, the company says. On 750MW that is at most 37.5MW, by our arithmetic. Cerebras says it has received no cancellations and expects to be back on schedule in the second half of 2027.
Business Pill · Backlog
A short explainer on backlog: revenue that is signed but not yet earned. It teaches the general idea only and does not discuss Cerebras.
The key insight: As we read the 10-Q and the call, the OpenAI delay is a timing problem with a price attached: credits accrue on late tranches from 1 October, and less than 5% of the 750MW, at most 37.5MW by our arithmetic, could become cancellable this year.
Feldman put the delay down to data centers, not chips. “After signing the agreement with OpenAI, it took us longer than we would have liked to find contracts for and bring online the substantial increase in data center capacity required to meet the aggressive requirements in the agreement,” he said. Asked whether land, power and building shells were the limit, he answered: “The answer is absolutely.” He said TSMC had committed enough wafers for 2027.
The 10-Q puts it in writing: as of 30 September, “we delivered less capacity to OpenAI than the quantity contemplated by the MRA”, the master relationship agreement signed in December 2025. It adds that “capacity delivery delays may continue through a portion of 2027”. Neither statement appears in the second-quarter 10-Q; it has no shortfall disclosure and no terminable-capacity estimate, so the disclosure is new this quarter, by our reading of both filings.
OpenAI capacity under its Cerebras contract: the 750MW commitment, the 1.25GW option, and the under-5% of committed capacity Cerebras says could be terminable in 2026. The 37.5MW is our arithmetic.
How the Delay Credits Work
The 10-Q sets out the mechanics. A shortfall that is not excused and remains uncured 30 days after a milestone date causes service credits to accrue on that tranche, and “such credits offset the monthly service fees payable by OpenAI to us in subsequent months”. If the shortfall remains uncured 90 days after the milestone, “we and OpenAI can each elect to terminate such tranche”. Delays also shorten the term of that tranche day for day.
Chief financial officer Bob Komin said Cerebras has “calculated the amount for potential delay credits and deducted them from RPO at September 30th”. He did not give the amount, saying only that the credits are “estimated to be relatively small” as a share of the backlog. Feldman called the figure “a very conservative number of what is the absolute worst case we could be hit with”. Komin said Cerebras had not assumed any of the delays would be excused.
The Backlog Before and After
Remaining performance obligations, the contracted revenue not yet recognised, were $26.1 billion at 30 September, against $25.4 billion at 30 June in the second-quarter 10-Q. So the backlog still rose $0.7 billion after the delay credits were taken out, by our arithmetic. Komin said OpenAI “represents the majority of our reported RPO to date”. Our results piece has the raised outlook against that backlog.
The cancellable share is where the money sits. If the backlog were spread evenly per megawatt, under 5% of OpenAI’s share would be worth under $0.65 billion if that share were just half the backlog ($13.05 billion), and could not exceed $1.3 billion, 5% of the whole $26.1 billion, by our arithmetic. That is a ceiling, not a forecast: Cerebras gives no dollar value per megawatt. If a tranche is terminated, the 10-Q says, the contract value and backlog “will decrease accordingly at such time”.
What OpenAI Pays Now, and the Warrant
OpenAI was 34% of third-quarter revenue, the 10-Q says, and Cerebras recognised $72.4 million under the arrangement, net of $3.2 million of warrant amortization. The warrant for up to 33,445,026 Class N shares vests in tranches tied to milestones that include “delivery of the Committed Capacity and Additional Capacity in tranches”, so a later delivery also means later vesting. Our piece on the customer warrants covers what OpenAI holds; the Microsoft deal is the other side of the same call.
Komin said Cerebras would resell any cancelled capacity: “if OpenAI or Cerebras were to elect an option to cancel a delayed capacity tranche, we believe there is sufficient demand and we would find other customers to absorb it”. Feldman said OpenAI has the right under the contract to take hardware instead, but that “our understanding from them is their data centers are also behind schedule”.
The Structural Read
As we read the two 10-Qs, the backlog absorbed the estimated credits and still grew, from $25.4 billion at 30 June to $26.1 billion at 30 September, so the delay shows up as a deduction Cerebras has not quantified rather than a fall.
As we read Feldman’s answers, the constraint he named is land, power and building shells, not chips, and he said OpenAI’s own data centers are also behind schedule.
Cerebras Systems, Form 10-Q for the quarter ended 30 September 2026
“we delivered less capacity to OpenAI than the quantity contemplated by the MRA”
Three Implications
FOR INVESTORS Komin said future changes to the delay-credit estimate will be reflected in both the backlog and guidance each quarter.
FOR OPENAI The 10-Q says either side can elect to terminate a tranche that remains uncured 90 days after its milestone.
FOR OUR COVERAGE We will report the delay-credit amount and any terminated tranche when Cerebras or OpenAI discloses them.
The framework says: “But infrastructure reality is catching up with AI ambitions: a projected 19GW gap between new data center demand and available grid capacity threatens to deflate the AI bubble.”
As we read the call, Cerebras’ delay fits that pattern: Feldman named land, power and building shells, not chips, as the limit on delivering OpenAI’s 750MW. The framework predates these filings and does not discuss Cerebras.
Cerebras’ remaining performance obligations at 30 June and 30 September 2026 from its 10-Qs, and the value of 5% of the backlog under an even per-megawatt spread. The 5% rows and the change are our arithmetic.
What Is Not Established
Cerebras has not disclosed the dollar amount of the delay credits, how many megawatts it has delivered, or which tranches are late. The 37.5MW and the dollar ceilings are our arithmetic on the company’s “less than 5%” and assume value spread evenly per megawatt, which the company does not say. OpenAI has not commented on the delay in anything we found; we did not contact either company. Whether any delay is later excused, which would cut the credits, is not known.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Cerebras has told investors it is behind on the first tranches of OpenAI’s 750MW contract, that service credits started accruing on 1 October, and that less than 5% of the committed capacity, at most 37.5MW by our arithmetic, could become cancellable this year. The backlog still rose to $26.1 billion after the estimated credits, and the company expects to be back on schedule in the second half of 2027.
A note on sourcing. We read Cerebras’ Form 10-Q for the quarter ended 30 September (filed with the SEC on 12 November) for the capacity shortfall, service-credit and termination terms, backlog, OpenAIrevenue and warrant; its 10-Q for the quarter ended 30 June for the earlier backlog; and The Motley Fool’s transcript of the 12 November earnings call for Feldman’s and Komin’s remarks. The megawatt and dollar ceilings and the backlog change are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
